Framework for spectral-spatial classification schemes

By using only spectral information in the classification of remote sensing images, we are not taking advantage of the relationship among adjacent pixels. Pixels from homogeneous regions are similar in their spectral response, indicating a relationship between them. It has been clearly stated that the spatial information extracted from the hyperspectral images improves the accuracy of the classification when it is incorporated into the scheme [15, 54, 154, 155, 41, 19, 29, 56, 57].

Figure 3.2 shows one possible spectral-spatial framework for classification of hyperspec- tral images that will be used as a base for our proposals. The spatial processing is mainly derived from techniques designed for greyscale and most of the commonly used methods are not appropriate forn-dimensional images. In the case of color and multispectral images, the visible (red, green, and blue) and infra-red bands can be processed independently. For hyper- spectral images, one way of applying spatial techniques is by reducing the number of spectral features to only a few or even one, for example, by using the first principal component com- puted by PCA or ICA. Unlike the spectral schemes, the feature reduction stage illustrated in Figure 3.2 is also applied as a first step to reduce the spectral dimensionality in order to extract the spatial information. Therefore, the techniques for reducing the dimensionality of hyperspectral images prior to the pixel-wise classification are also of special interest in the spectral-spatial classification schemes.

Different methods to reduce the number of features on the generation of spectral-spatial classification schemes were investigated in [155, 31]. Multidimensional and vectorial gradient methods, such as the RCMG described in Section 2.2.2, were used in [155] to reduce the number of bands to one. In [31] the research focused on the best techniques for reducing the number of features to create schemes based on mathematical morphology, concluding that other methods than PCA may be more adequate in terms of classification accuracy for creating those schemes.

Additional preprocessing may be applied as a first step, as illustrated in Figure 3.2, for improving hyperspectral image classification [164, 166], for example by denoising the hyper- spectral data.

3.1. Introduction 63

Figure 3.2:Framework of a spectral-spatial classification scheme. Some schemes require a postprocessing for joining the spectral and spatial results.

The spatial information is extracted from the closest neighborhood of a pixel. The closest neighborhood can be defined by a fixedn×nwindow, such as the structuring element used in MM. The neighborhood can also be defined by a homogeneous region, named adaptive neigh- borhood as the number of neighboring pixels depends on the size of the region considered. The regions created in a segmentation map, or the flat zones created by a self-complementary area filter [56] are two examples of adaptive neighborhood.

A method for extracting spatial features based on opening and closing by reconstruction, see Section 2.5.1, was proposed in [124]. In order to apply the morphological approach to hyperspectral images, PCA is first applied on the hyperspectral data, and the most significant PCs are used as base images, creating a new pixel vector of morphological features, known as EMP [15]. The spectral information in the EMP is small compared to the spatial information incorporated by the morphological profiles. Fauvel et. al [54] proposed a scheme for urban land cover classification of hyperspectral images, which extended the EMP by fusing spectral data. The spectral information is incorporated from the original hyperspectral data without any additional processing.

Regarding the segmentation, the spectral-spatial classification scheme presented in [158] combines the results of a pixel-wise SVM classification and a segmentation map. Several seg- mentation techniques have been proposed for this scheme, such as partitional clustering [154], watershed transform [155, 137] and hierarchical segmentation (HSEG) [160]. Section 2.3 has a summary of these techniques. Clustering algorithms were used in [154], and evolutionary CA were designed and applied in [132] for segmenting hyperspectral images, although in this last case reducing the dimensionality of the data is not required. The idea behind these schemes is to regularize the thematic map produced by the classifier by using the regions of the segmentation map as an adaptive neighborhood.

The classification result by spectral-spatial schemes, as illustrated in the final classifica- tion map shown in Figure 3.2, gives rise to greater accuracy and more homogeneous thematic maps as compared to spectral classification schemes. In order to increase the spatial infor- mation included in the schemes, a spatial post-regularization (PR) can be applied over the thematic map produced by the classifier [154]. This step that is applied in [154] included as part of the post-processing in Figure 3.2, introduces additional spatial/textural information by considering separately for each pixel the classes that the classifier assigns to the neighbors in a fixedn×nwindow. The new class for each pixel is decided by analyzing information of other pixels in a fixed-size neighborhood. This regularization is performed until stability is reached (none of the pixels changes its class) [158]. By repeating the regularization process until stabilization of the class labels, the regions in the final thematic map are homogeneous. Thus, the spatial PR can be applied to the thematic map before a MV or to the final thematic map produced by the spectral-spatial classification scheme.